Security thesis

Keep the castle. Distribute the gold.

Your customer data lives in a dozen systems that don’t talk to each other, and every team builds a taller wall around its own island. The safest-sounding answer — put it all in one big vault — is the one that will hurt you most.

The incumbent’s mistake

It isn’t having walls. It’s hoarding all the gold in one vault behind them.

That vault is the blast radius: one breach takes everything. And while it sits there, the hoarding also locks the value away from the people who need it. A single store of everything is simultaneously the biggest prize and the tightest bottleneck.

The fortress is the breach.

Keep the castle. Keep the moat.

None of this means tearing down the perimeter. Cloudflare’s walls still stand: the firewall, the DDoS shield, the edge governance. Keep every one of them.

What changes is what being inside them earns you. Nothing. Every request is re-verified, per action, whether it arrives from the open internet or from a service that has been running inside the walls for a year.

The castle governs access. It doesn’t store the treasure.

Distribute the gold

Instead of one vault, the value sits in many pockets — each in the system genuinely authoritative for it:

Breach one and you get a fragment, never the hoard. There is no single secret to steal and no single point to take down.

And because the value is spread out for people and agents to use, distributing it is the accessibility. Security and reach in the same move — which is normally the trade you are asked to make.

Substrate is not security

People hear no-code and assume toy. But the tools are only where this runs — not what makes it safe. The controls are the primitives enterprise systems use: signed mandates, PKCE and OIDC, per-action authorization, offline-verifiable grants, real-time consent.

Judge the boundary logic, not the vendor logos.

The same core capability, operated by the person who needs it

The clearest way to say what this is: a Vault-grade key service that a designer or business analyst can operate. Not a lighter imitation of one — the same primitives, without the cluster.

Where Vault is still the right answer: dynamic database credentials, acting as a private PKI certificate authority, SSH certificate signing, and the breadth of its secret-engine ecosystem. If you need those, run Vault — the two are not mutually exclusive, and the walls are the same walls.

What Vault was never built for is the part that matters once agents start acting: an artifact a customer can be granted, a signature a third party can verify without asking you, and a permission that follows a person rather than a machine account. Vault predates that problem. It was designed when the actors were services you deployed, not agents transacting on someone’s behalf.

Why a closed boundary fails an AI actor

A closed, single-vendor system governs by containment. That is fine at human speed. An AI actor moves faster than a sealed system can respond, and containment cannot do the three things AI governance needs — while doing a fourth nobody wants.

You hold the keys

The keys stay in your hand — owned, not rented from a platform. The proof of trust is a public key, so there is nothing secret for anyone to take, and nothing you have to ask us for in order to verify.

Keep the castle. Distribute the gold. Hold the keys. Nothing concentrated to steal, and everything reachable to the people — and the agents — you have allowed.